China’s large model has a weekly call volume of 33 trillion Tokens: No. 1 in the world for 13 consecutive weeks, 14 times that of the United States

What is the concept of 33 trillion Tokens? Assuming that a Chinese article has about 1,000 words and about 1,500 Tokens, this number is equivalent to the amount of about 22 billion articles processed by China's large model in a week. On July 29, the latest data disclosed by the model aggregation platform OpenRouter showed that the weekly call volume of China's large AI models reached 33 trillion Token last week, 14.1 times that of the United States during the same period, and has ranked first in the world for consecutive weeks.

A quantitative profile

This set of data outlines the explosive growth of China's large model application demand from a quantitative perspective. The weekly call scale of 33 trillion Tokens means that the penetration depth and activity of domestic large models in various application scenarios such as dialogue, coding, content generation, and agents have reached the world's leading level, and there is a clear gap between the United States and the second place. It is not an advantage of a few percentage points, but a generation difference in orders of magnitude.

Behind the continued leadership in call volume is the joint action of both supply and demand. On the supply side, the maturity of large-scale domestic models has increased rapidly - the blowout of open source models, continued price drops, and expansion of inference infrastructure have made "affordable and accessible" a reality; on the demand side, from DeepSeek to Kimi, Qwen, the user base and application scenarios of the top domestic models are expanding simultaneously.

The other side of leading

From an industry perspective, this set of data on the one hand confirms the large-scale implementation capability of China's AI application ecosystem - on a global scale, the only economy in the world that can achieve this level of model calls is China. But on the other hand, it also raises a more sober question: Has the lead in the number of calls been simultaneously transformed into a closed loop of high-quality application value and commercialization? In the context of price wars and free strategies, it is still an open question whether the scale effect of 33 trillion Tokens can cover the cost of computing power.

What deserves more attention is the structure of the call volume - whether it is a large number of low-value, low-viscosity tentative calls or high-value core business calls, which determines the quality of this set of data. For domestic large model manufacturers, moving from "traffic leadership" to "value leadership" is a key question in the second half of 2026.

Several directions worth tracking in the future:

1. Structural Analysis of Call Volume: The proportion of dialogue, coding, Agent and other scenarios can explain the problem better than the total volume.
2. Sustainability of the lead: The speed at which U.S. manufacturers catch up will test whether the 13-week lead is a trend or a peak.
3. Commercial conversion rate: Whether the scale effect can cover the cost of inference determines the commercial value of leading calls.
4. The relationship between price war and calls: Is there any "false fire" component in the increase in call volume driven by price cuts?